Undiv io

Wan 2.2 Image to Video (I2V) - Lightx2v Enhanced Motions

CREDITS: https://civitai.com/models/1905937/wan-22-14b-i2v-t2v-lightx2v-enhanced-motions?modelVersionId=2157276 Workflow Overview & Strategy This isn't just a standard pipeline; it's a carefully engineered process: Image Preparation: The input image is automatically scaled to the optimal resolution for the Wan model. Dual-Model Power: The workflow leverages both the Wan 2.2 High Noise and Low Noise models, patched for performance (Sage Attention, FP16 accumulation). The "Secret Sauce" - LoRA Overclocking: The Lightx2v LoRA is applied at significantly elevated strengths: High Noise UNet: 5.6 (The primary driver for introducing strong motion) Low Noise UNet: 2.0 (Refines the motion and cleans up the details) Staged Sampling (CFG++): A two-stage KSampler process: Stage 1 (High Noise): 4 steps to generate the core motion and structure. Stage 2 (Low Noise): 2 steps to refine and polish the output. (Total: 6 steps). Post-Processing: The generated video sequence is then upscaled with RealESRGAN and the frame rate is doubled using FILM interpolation for a buttery-smooth final result. Tips & Tricks Prompt is Key: For the best motion, use strong action verbs in your positive prompt (e.g., "surfs smoothly," "spins quickly," "explodes dynamically"). Experiment: The LoRA strengths (5.6 and 2.0) are my tested "sweet spot." Feel free to adjust them slightly (e.g., 5.4 - 5.8 on High Noise) to fine-tune the motion intensity for your specific image. Resolution: The input image is scaled to ~0.25 Megapixels by default for speed. For higher quality, you can increase the megapixels value in the ImageScaleToTotalPixels node, but expect longer generation times.

API
Wan
Video & Animation
Image to Video